This Week in AI: November 24-30, 2025 (Week 48)

News of the week: November 24-30, 2025 (Week 48)

The week opened with Anthropic releasing Claude Opus 4.5, a model that raised the bar for coding while also getting much cheaper. The chip market was stirred by a report that Meta is in talks to buy TPU chips from Google, prompting a public response from Nvidia. OpenAI focused on shopping ahead of Black Friday, but also had to deal with a security incident at a vendor and a lawsuit over a teenager’s suicide. In the US, the government launched the Genesis Mission initiative, and Amazon pledged up to 50 billion dollars for AI infrastructure for US agencies.

New Models

Anthropic Releases Claude Opus 4.5

On November 24, Anthropic introduced its most powerful model yet, Claude Opus 4.5. The company focused it on coding, agents (AI that carries out multi-step tasks on its own) and computer use. On SWE-bench Verified, a test measuring the ability to fix real software bugs, it reportedly scored 80.9%. (MacRumors)

The big change is the price: 5 dollars per million input and 25 dollars per million output tokens (tokens are the chunks of text AI usage is billed by). That is roughly two thirds less than its predecessor, Opus 4.1. Alongside the model, Anthropic expanded the Claude for Chrome extension to Max subscribers and Claude for Excel to business plans. (AlternativeTo)

Anthropic also calls Opus 4.5 its safest model to date. It is better at resisting prompt injection, meaning attempts to slip the AI malicious instructions hidden, for example, in a web page. (SiliconANGLE)

DeepSeekMath-V2: An Open Model at Math Olympiad Level

On November 27, China’s DeepSeek released DeepSeekMath-V2, a model with open weights that anyone can download. According to the company, it achieves results equivalent to a gold medal at the 2025 International Mathematical Olympiad and scored 118 out of 120 points on the 2024 Putnam competition. (Simon Willison)

What makes it interesting is its self-verification approach: one part of the model writes mathematical proofs and another, the so-called verifier, checks them. The model has roughly 685 billion parameters and is built on top of DeepSeek-V3.2.

Black Forest Labs Launches FLUX.2

On November 25, German company Black Forest Labs released FLUX.2, a family of image models that can both generate and edit images. They can work with up to ten reference images, render text better and handle edits at around 4 megapixels. (TestingCatalog)

Part of the family is available through an API, part with open weights. The company worked with Nvidia and the ComfyUI project on optimizing it for graphics cards. (AlternativeTo)

Microsoft Fara-7B: A Small Agent That Operates a Computer

On November 24, Microsoft Research released Fara-7B, a model with seven billion parameters. It looks at screenshots and clicks and types based on them, which lets it operate websites. It is small enough to run directly on the user’s computer without sending data to the cloud. (Microsoft Research)

On the WebVoyager benchmark it scored 73.5%, more than GPT-4o in the same setup.
Before sensitive steps such as a payment or entering personal data, it stops on its own and waits for a human.
It is available under the MIT license on Hugging Face; Microsoft describes it as experimental.

Products and Agents

ChatGPT Helps With Shopping and Talks Right in the Chat

On November 24, OpenAI launched a shopping research feature. The user describes what they are looking for, and within a few minutes ChatGPT puts together a clear buyer’s guide comparing products. The feature is available on the free plan too, with nearly unlimited use over the holidays. (OpenAI)

The next day, November 25, OpenAI moved voice mode directly into the regular chat window. During a conversation, answers appear as text, and ChatGPT can also show maps or the weather. The original separate window can be turned on in settings. (MacRumors)

AI Shopping in Numbers

November 25: PayPal and Perplexity launched Instant Buy in the US, which lets users pay for purchases directly in a conversation with AI. (PayPal)
November 29: According to Adobe Analytics, Americans spent a record 11.8 billion dollars online on Black Friday, 9.1% more than last year. (Nasdaq)
Traffic to retail sites coming from AI assistants grew 805% year over year between November 1 and 28. (The Star / Reuters)

Alibaba Bets on Consumers

November 24: Alibaba‘s Qwen app passed 10 million downloads in the first week of its public beta, according to the company; its shares in Hong Kong then rose more than 5%. (TechRepublic)
November 27: Alibaba began selling Quark AI Glasses smart glasses in China, powered by Qwen models; they can, for example, translate or take meeting notes. Prices start at 1,899 yuan. (Engadget)

Business and Investment

Meta in Talks Over Google Chips, Nvidia Pushes Back

The Information reported that Meta is in talks to buy TPU chips from Google worth billions of dollars. TPUs are processors Google designed specifically for AI. Meta could deploy them in its own data centers from 2027 and start renting them through Google Cloud even earlier. (The Register)

Nvidia shares fell as much as 7% intraday on November 25 and closed about 2.6% lower, while Alphabet shares rose. (Euronews)

Nvidia responded on X the same day, saying it was “delighted by Google’s success” but that it is “a generation ahead of the industry.” It stressed that its graphics processors are more versatile than single-purpose ASIC chips, the category TPUs belong to. (Tom’s Hardware)

Amazon: Up to 50 Billion Dollars on AI for the US Government

On November 24, Amazon announced it will invest up to 50 billion dollars in AI and supercomputing for AWS government customers. Construction is set to begin in 2026 and add roughly 1.3 gigawatts of capacity in regions for classified and top secret government workloads. (BNN Bloomberg / Reuters)

Agencies will get access to the Bedrock and SageMaker services, to Amazon Nova and Claude models, and to Trainium and Nvidia chips.

Earnings Show the Boom and Its Cost

November 25: Alibaba reported 34% growth in cloud division revenue to 39.8 billion yuan, but profit fell sharply due to AI investments. (AP / Denver Gazette)
November 25: Dell took in record quarterly orders for AI servers worth 12.3 billion dollars and raised its full-year outlook. (Dell)
November 25: HP will cut 4,000 to 6,000 jobs by the end of fiscal 2028 as part of a restructuring tied to adopting AI; it aims to save about a billion dollars a year. (TechRepublic)

Regulation and Law

Trump Launches the Genesis Mission

On November 24, President Donald Trump signed the Genesis Mission executive order. It tasks the Department of Energy and the national laboratories with building a digital platform that brings scientific data together and uses AI to speed up research. (ABC7 / AP)

Within 60 days, the department must identify at least 20 science and technology challenges for the initiative to focus on. The order is often compared to the Manhattan Project. (Maynard Nexsen)

OpenAI Denies Responsibility for Adam Raine’s Death

In its first court filing in the case brought by the family of Adam Raine, a 16-year-old who died by suicide, OpenAI argued on November 25 that ChatGPT did not cause his death. The company says the chatbot urged him to seek help more than a hundred times and that he violated its terms of use. (NBC News)

The family’s lawyer, Jay Edelson, called the response “disturbing.” The case is one of several lawsuits raising the question of AI companies’ responsibility for users’ mental health.

Warner Music Reaches a Deal With Suno

On November 25, record label Warner Music Group settled its copyright dispute with song generator Suno and entered into a partnership with it. In 2026, Suno will release new licensed models and retire its current ones. (Music Business Worldwide)

Artists will decide whether and how their name, voice or songs are used.
Song downloads will be limited to paying users, with monthly caps.
Suno also bought the Songkick platform from Warner; disputes with Universal and Sony are ongoing.

Research

Anthropic Estimates AI’s Productivity Gains

On November 25, Anthropic published a study based on an anonymized analysis of 100,000 conversations with Claude. By its estimates, tasks would take about 90 minutes on average without AI, and Claude cuts that time by roughly 80%. (Anthropic)

With widespread adoption, this could raise US labor productivity growth by 1.8 percentage points a year. The authors themselves caution that this is not a forecast and that the figures may overstate the real benefit, since they do not account for work done outside the conversation.

Sutskever: The Age of Scaling Is Ending

OpenAI co-founder and head of Safe Superintelligence, Ilya Sutskever, gave an interview to Dwarkesh Patel‘s podcast on November 25. He argues that the 2020-2025 period, when progress came mainly from making models and compute bigger (scaling), is ending and a new “age of research” is beginning. (The Neuron)

Security and Incidents

Data Breach at an OpenAI Vendor

On November 26, OpenAI disclosed a security incident at Mixpanel, an analytics service it used on its developer website. An attacker exported data on some API users from Mixpanel’s systems, such as names, email addresses and approximate location. (OpenAI)

OpenAI’s own systems were not breached, and the attacker did not get access to passwords, API keys, conversations or payment details. OpenAI stopped using Mixpanel and advised users to watch out for phishing and to turn on multi-factor authentication. (SiliconANGLE)